itdxer

There are a couple of advantages:

1. It's based on Theano, so it's fast. Also you are able to run your code on CPU and GPU.

2. NeuPy supports a lot of algorithms and different layer types (http://neupy.com/docs/cheatsheet.html), so you can easily construct deep neural networks.

3. In the NeuPy it's easy to read and understand network's structure. One of the newest features is a Subnetworks (http://neupy.com/docs/layers/basics.html#subnetworks).

4. Neural Network Surgery is another new feature (http://neupy.com/docs/surgery.html)

Usually each library has it's own architecture. This architecture based on some default classes that implements core functionality for models. I already have a chance to check source code for the popular neural network libraries and I found that their architecture requires some basic classes that are related to the Backpropagation algorithm and it's variation. In neural networks there are a lot of algorithms that really hard to compare and generalize. To make your idea possible I see two ways for that. First one is to change architecture for the main models classes in popular library. In practice for the popular libraries it's a huge amount of work. Even if you found possibility to generalize core architecture with other network that you want implement, probably, you will broke a big part of library that works before. The second one is build a new folder inside of existing library that contains localized code that don't have relation to the core part. How it differs from the decision of a single repository for a different library? In practice its a good way to separate something without relation, because further attempts to combine all in a big library will lead to breakage and bugs that are difficult to debug and fix.

There are two main advantage of ASCII approach. The first one is that vectors weight is not very big and you can build relatively small matrix that easy to store and faster to use. The second one is that usual encoding alternatives give a vary sparse vectors and as a consequence there would a big number of negative values (because we encode 0 as -1 value) inside the weight. For instance, if you have something encoded as 1 (word, letter or anything like that) with a big number of zeros (which would be encoded as -1 value in Discrete Hopfield network) that would be very unlikely to see the same ones at the same place multiple number of times and after some number of iterations negative values will reduce value encoded as 1 inside of memory and these -1 values would be as a dominant value for specific pattern.

Thank you for your response. There are would be more interesting projects related to neural networks.

And also thank you for detecting this problem with [source] link. I fixed it.

It's an interesting idea. But I guess there are would be two main problems.

1. Memory is limited. For this task you need to save a lot of words to make possible use it in applications. To solve this problem you should save words in a huge dimensional vectors.

2. This network can't solve the problem of missing letters in words with a high accuracy.

I know that there are some great projects related to neural networks. But they are focused on deep neural networks. Apart from this trend in neural networks there are many other great ideas like Autoassociative memories, learning base on associations, Adaptive resonance theory and so on. They are really cool and you can make a lot of interesting projects with them. This was a major inspiration for this library.